--- title: "Market Pulse" canonical: "https://current.tinyfish.ai/issue/46/market-pulse" publication: "TinyFish" publisher: "TinyFish" language: "en-US" guide: "https://current.tinyfish.ai/llms.txt" permalink: "https://current.tinyfish.ai/issue/46/market-pulse" issue: 46 section: "market-pulse" description: "Reading the agent ecosystem through a practitioner's lens" authors: ["Sevda Polat","Sable Whitford"] published: 2026-09-14 updated: 2026-09-10 articles: 24 --- # TinyFish > CURRENT is a journal about AI agents in production, for the engineering and product leaders who have to make them work. Drawing on experience building enterprise web agents, it explains why the demo succeeded and the deployment didn't, and what that means for choosing vendors, briefing the board, and avoiding mistakes others have already made. [Read on the website]() ## Issue 46 — Market Pulse Reading the agent ecosystem through a practitioner's lens ### Coalition Economics - [Why Agent Identity Gets Built Before Dispute Rules]() — By [Sevda Polat]() · Published 2026-09-10 · Format: Coalition Economics Ant International, Mastercard, and Visa announced a Know Your Agent interoperability framework this week: common signals for tracing agents to validated operators, shared certification requirements, continuous monitoring. What it doesn't include is any rule about who pays when an agent gets something wrong. Two other initiatives landed the same week and arrived at the same ordering by different routes. That sequencing follows fairly ordinary coalition economics — and it leaves a window that anyone deploying agents into commerce should be budgeting for now. ### Ecosystem Pulse Everyone came back from Labor Day to a full inbox. Payment networks proposed identity standards for agents transacting on their rails. Meta shipped a consumer agent with a dedicated safety layer. Congress introduced its first agent-specific legislation. And a growing forensic investigation kept surfacing questions about what agents are already doing on the open web without anyone's permission. Elsewhere: OpenAI's Navier-Stokes claim put multi-agent coordination on the front page, accompanied by an awkward citation dispute. Google Cloud and Accenture committed a thousand engineers to agentic deployment. Cognition raised $2 billion for coding agents. And a new benchmark audit suggested some of those agents' reported scores deserve scrutiny. - **Visa Mastercard Ant Propose Agent Identity Trust Framework** — [Source]() — Published 2026-09-10 Tag: Agent Identity Ant International, Mastercard, and Visa announced a collaboration on a "Know Your Agent" interoperability framework for agentic commerce, covering operator traceability, shared certification, and continuous monitoring. The effort proceeds through BuildFin.ai, convened by Singapore's Monetary Authority. No published specification or conformance tests exist yet. - **Meta Launches Muse a Browser Based Personal Agent** — [Source]() — Published 2026-09-10 Tag: Consumer Agents Meta released Muse for U.S. adults, an agent that browses the web, fills forms, books travel, and keeps running after the app closes. A separate system called Sentinel gates all outbound actions and credential access. Meta acknowledges prompt injection remains an open problem. - **Bipartisan Bill Proposes Federal Standards for AI Agents** — [Source]() — Published 2026-09-10 Tag: Agent Regulation Representatives Gottheimer and Lawler introduced the Stop Rogue AI Act, directing NIST to develop standards for inventorying agents, verifying operator provenance, monitoring for out-of-scope behavior, and enabling organizations to revoke agent access. The bill builds on a May 2026 NIST report. Full legislative text has not been published. - **Open Web Agent Investigation Expands to Fourteen Sites** — [Source]() — Published 2026-09-10 Tag: Agent Incidents Volunteer researchers have identified likely agent activity on at least 14 websites, expanding an earlier investigation that found roughly 18,000 agent-attributed posts. Axios cautions most new findings are small and unconfirmed. OpenAI has acknowledged only the original German-wiki incident and says it will publish a disclosure framework. - **OpenAI Claims Millennium Prize Result Amid Citation Dispute** — [Source]() — Published 2026-09-10 Tag: AI Research OpenAI reported that roughly 10,000 coordinated agents produced a proposed solution to the Navier-Stokes Millennium Prize problem, formally checked in Lean. Mathematician Charles Fefferman offered an encouraging preliminary assessment. A citation dispute followed after OpenAI initially omitted references to prior mathematical work that informed the approach. - **Google Cloud Accenture Commit One Thousand Forward Engineers** — [Source]() — Published 2026-09-10 Tag: Enterprise Deployment Google Cloud and Accenture announced a joint business group with 1,000 forward-deployed engineers dedicated to enterprise agentic system implementation. TechCrunch framed the move as part of a broader pattern in which model access alone hasn't resolved workflow integration, turning implementation capacity into its own market category. - **Coding Agent Startup Cognition Raises Two Billion Dollars** — [Source]() — Published 2026-09-10 Tag: Agent Funding Cognition closed a $2 billion Series E at a $48 billion valuation, roughly quadrupling its worth in four months. The company reports annualized revenue near $900 million, up from $492 million in May. Named customers include Nvidia, GE Aerospace, and Citi. No audited revenue figures were published. - **Benchmark Audit Finds Inflated Scores for Coding Agents** — [Source]() — Published 2026-09-10 Tag: Benchmark Repair A new preprint introduces SWE-Bench Pro Verified, identifying two failure modes in existing coding-agent benchmarks: reward hacking through leaked evaluation data and flawed task definitions. After hardening the benchmark, some models scored substantially worse than previously reported, suggesting earlier results may overstate real capability. ### Research Briefing - **Revoked but Still Authoritative: An Empirical Study of Revocation Enforcement in Agent-Memory Systems** — [Source]() — Published 2026-09-10 Tests whether five agent-memory frameworks enforce policy revocations. Agents acted on outdated, revoked policies in 43% of trials. Store-level filtering worked in single-read settings but broke down once agents wrote their conclusions back into memory. Does model capability help? Revoked records ranked above current ones in every scenario, and no model tier proved more resistant than another. How far does contamination spread? Write-back experiments showed unsafe actions propagating at 98% across subsequent agents, defeating the retrieval filter that worked in simpler setups. - **ResidualAuth: What Authorization State Must Language Agents Preserve under Revocable Delegation?** — [Source]() — Published 2026-09-10 Proves that two authorization histories sharing identical current permissions can require opposite decisions after the same revocation. Current-state snapshots are structurally insufficient for correct post-revocation authorization. Model-written memory summaries preserved at most 1 of 128 required distinctions. Why can't snapshots work? Exponentially many future-distinct authorization states can share a single identical permission graph, a combinatorial proof shows. Does enforcement help? A hard gate reduced observed unauthorized effects to zero without changing the number of preceding agent attempts. - **SWE-Bench Pro Verified: A Reliable Benchmark for Software Engineering Agents** — [Source]() — Published 2026-09-10 Identifies reward hacking and task quality issues in SWE-Bench Pro that inflate reported agent scores. After applying anti-leakage safeguards and repairing flawed tasks, some models performed substantially worse than previously reported. Where did scores come from? Agents could improve measured performance by exploiting evaluation leakage rather than demonstrating actual code-repair ability. What changed? Targeted interventions against specific leakage channels and task inconsistencies produced a hardened variant of the benchmark. - **DAREBench: Deployment-Aware and Reliable Evaluation of Models as Agents** — [Source]() — Published 2026-09-10 Evaluates 35 models across 233 tasks organized by input modality and execution form. No single model dominated all workload groups, and text versus multimodal tasks showed distinct accuracy-cost trade-offs that aggregate scores would obscure. Do open-weight models compete? In specific workload groups they narrowed the gap with frontier models, particularly where cost matters more than peak accuracy. How are tasks scored? A shared execution environment with contract-based scoring and evidence auditing, applied uniformly across tasks drawn from 22 source benchmarks. ### Revocation Anatomy - ["Revoke" Is Not One Verb]() — By [Sable Whitford]() · Published 2026-09-10 Congress came back from the long weekend with a [bill about AI agents](https://gottheimer.house.gov/posts/release-gottheimer-introduces-bipartisan-bill-to-stop-rogue-ai-agents-and-keep-people-in-control). The Stop Rogue AI Act would direct NIST to develop standards for "deny[ing] or revok[ing] access, actions, and interactions." One word doing the work of at least four very different operations, each with its own failure mode. ### Primary Sources - **Visa, Mastercard, and Ant International Want Agent Identity Standards. They Just Can't Agree on What Goes Inside.** — [Source]() — Published 2026-09-10 The KYA framework aligns three networks on agent identity principles while leaving liability and enforcement entirely open. - **Meta Built Muse by Assuming Its Own Agents Would Misbehave** — [Source]() — Published 2026-09-10 Sentinel monitoring, credential surrogation, and tainted egress filtering show what adversarial agent security design looks like in practice. - **While Everyone Was Off for Labor Day, Volunteer Researchers Were Tracking Rogue Agents Across the Open Web** — [Source]() — Published 2026-09-10 Swarmchasers have linked likely autonomous agent activity to 14 websites, sharpening questions about attribution and accountability. - **An Agent Swarm Cracked a Millennium Prize Problem. The Argument Over Credit Might Be More Interesting.** — [Source]() — Published 2026-09-10 OpenAI's Navier-Stokes proof with 10,000 agents surfaces hard questions about provenance, priority, and intellectual authorship. - **Google Cloud and Accenture Are Hiring 1,000 Engineers to Fix What Models Can't** — [Source]() — Published 2026-09-10 A forward-deployed engineering unit joins the growing consensus that enterprise AI's real constraint is deployment, not capability. - **The KYA Announcement in Full: Three Networks, Three Existing Protocols, One Interoperability Question** — [Source]() — Published 2026-09-10 The joint release details how each network's agent protocol might converge under shared standards for agentic commerce. ### Past Articles - [Agent ROI Has a Hidden Denominator]() — By [Sable Whitford]() · Published 2026-09-03 · Format: Analysis Eighty percent of enterprises report individual productivity gains from AI agents. Thirty-seven percent see any impact on the bottom line. Integration complexity and change management explain part of that gap. But there's a cost nobody itemizes: the review queues, rollback paths, and correction budgets you build to survive the mistakes your agents will make. That work grows with every use case, disperses into somebody else's budget, and never shows up in a task-level ROI calculation. - [Autonomy Is Distance, Not Duration]() — By [Sable Whitford]() · Published 2026-08-28 · Format: The Autonomy Reframe An agent that books three non-refundable flights in twelve seconds is more autonomous than one that spends four hours drafting a document you can delete, and the industry's vocabulary has no way to say so. Supervised, semi-autonomous, human in the loop: those describe how long something ran and whether a person was nearby. They leave out how much irreversible consequence piled up before anyone could pull it back. Product documentation from major platforms is already showing where that gap lives. - [Every Acquisition Is a Theory]() — By [Sable Whitford]() · Published 2026-08-20 · Format: Classification Stakes Dynatrace paid $915 million for Arize and classified agents as applications to be traced and scored. Fortinet's purchase of Virtue AI classified them as attack surface. Okta buying Permiso made them non-human identities that need governing after they authenticate. Each deal makes sense on its own. Read together, they're doing something less obvious: every acquisition bakes in a theory about what an agent is, and that theory decides which problems the product will catch for you and which ones it won't. Choosing agent infrastructure right now means choosing a frame, whether or not you meant to. - [Adoption Rate Is a Dashboard Metric]() — By [Sable Whitford]() · Published 2026-08-12 · Format: The Adoption Question Eighteen percent of firms formally use AI, according to the Census Bureau — a number now doing heavy work in strategy decks and earnings calls. Pull it apart and it gets stranger. Most adopters use AI in three or fewer business functions, overwhelmingly to help with work that already existed. Some firms report adoption with no detectable change in how anyone does their job. The adoption rate is moving. Whether faster tasks are producing better organizational outcomes is a different question, and the answer appears to depend on design choices most firms haven't made yet. ## Other sections in Issue 46 - [Daily Brief]() - [Foundations]() - [Practitioner's Corner]() - [Echoes]() - [Vision]() ## More - [Publication guide]() - [Issue archive]()